e21068 Background: In patients with non-small cell lung cancer (NSCLC) treated with immune checkpoint inhibitors (ICPI), individual clinical, biological, and imaging prognostic biomarkers have recently been identified. However, the combination of these biomarkers has not yet been studied. This study aims to analyse various clinical, biological and 18FDG PET/CT parameters using machine-learning algorithms to build more accurate prognostic models of NSCLC response to ICPI. Methods: The exploratory cohort consisted of patients with metastatic NSCLC, treated with either pembrolizumab or nivolumab in monotherapy, included in two prospective cohort studies of our institution. For all patients, a total of 28 baseline quantitative features were collected just before the initiation of ICPI (12 clinical, 6 biological and 10 PET/CT parameters), such as the patient’s ECOG performance status, PD-L1 tumour expression level (PD-L1%), neutrophil to lymphocyte blood ratio, number and metabolism of lesions, metabolic tumour volume (MTV). The two endpoints were the progression-free survival at 6 months (6M-PFS) and the overall survival at 12 months (12M-OS). The 28 features were sorted according to their selection frequency by a LASSO logistic regression on extensive cross-validation to predict 6M-PFS and 12M-OS. The top eight features were selected using the intersection of the most frequently selected features of both outcomes and used to build a logistic regression to predict 6M-PFS and 12M-OS. An independent validation cohort was collected to assess the performance of the 8-features model. This cohort consisted of metastatic NSCLC patients treated with monotherapy ICPI and retrospectively included in three different hospitals: Nice (N =16), Monaco (N =10) and Rouen (N =19). We reported the area under the receiver operating characteristic (AUROC). Results: 117 patients were included in the exploratory cohort (N = 93 for training and features selection and N = 24 for the test set). Without any feature selection, the AUROC of the logistic regression was 73.6% and 77.6% on the test set. Using LASSO, the top eight selected features were: age, MTV, number of lesions, PD-L1%, smoking status, ECOG, SUVmax of the most intense lesion and blood sugar level. The AUROC of the 8-features model was 74.3% for 6M-PFS and 85.5% for 12M-OS on the test set. 45 patients were retrospectively included in the validation cohort. The AUROC of the model were 81.9% for the 6M-FPS and 88.2% for the 12M-OS. Conclusions: The combination of heterogeneous biomarkers provides a powerful model for predicting the outcomes of NSCLC patients treated with ICPI. Rigorous feature selection is a critical point in ML approaches to avoid overfitting. [Table: see text]